Deep Learning Provides Substantial Improvements to County‐Level Fire Weather Forecasting Over the Western United States

The recent wildfires in the western United States during 2018 and 2020 caused record‐breaking fire damage and casualties. Despite remarkable advances in fire modeling and weather forecasting, it remains challenging to anticipate catastrophic wildfire events and associated damage. One key missing component is a fire weather prediction system with sufficiently long lead time capable of providing useful regional details. Here, we develop a hybrid prediction model of wildfire danger called CFS with super resolution (CFS‐SR) as a proof of concept to fill that void. The CFS‐SR model is constructed by integrating the Climate Forecast System version 2 with a deep learning (DL) technique from Single Image Super Resolution, a method widely used in enhancing image resolution. We show that for the 2018–2019 fire season, the CFS‐SR model significantly improves accuracy in forecasting fire weather at lead times of up to 7 days with an enhanced spatial resolution up to 4 km. This level of high resolution provides county‐level fire weather forecast, making it more practical for allocating resources to mitigate wildfire danger. Our study demonstrates that a proper combination of ensemble climate predictions with DL techniques can boost predictability at finer spatial scales, increasing the utility of fire weather forecasts for practical applications. Plain Language Summary Fire weather prediction is significantly improved by a hybrid approach‐ a weather forecast model and machine learning algorithm. We train a deep learning (DL) technique with model simulations to directly produce a fire hazard estimate based solely on predicted weather conditions. The predictability of extreme danger fire conditions is enhanced which can be useful in reducing fire pre‐suppression activities. The applied DL technique also successfully produces a higher spatial resolution (4 km) product for county‐level fire forecasting. This approach shows that DL can be integrated in weather/climate models to produce better quality. Key Points Combining dynamical model prediction and deep learning (DL) can boost predictability of wildfire risk up to 14 days DL efficiently reduces the internal bias of dynamical model and enhances feature depiction up to 4 km resolution Predictions, especially associated with extreme fire weather, are well grounded on the characteristics of wind behavior